For decades, data center power architecture followed a fairly familiar pattern. Utility power entered the site as AC, was transformed, distributed, converted, protected, stepped down, converted again, and eventually delivered to servers that were ultimately running on DC inside the rack.

That system worked because the loads were manageable. A 5 kW rack was not unusual. A 10 kW rack was considered serious. Even as cloud facilities grew into massive campuses, the basic power chain could still be stretched with larger switchboards, larger UPS systems, better power distribution, and tighter operational discipline.

AI is changing that math.

The conversation is no longer about whether a rack consumes 10 kW or 20 kW. The industry is now talking about 100 kW racks, 300 kW racks, and eventually rack-scale systems approaching 1 MW. In a recent pv magazine interview, SolarEdge noted that most data centers today may be around 10 kW per compute rack, while the next generation entering AI data centers is closer to 100 kW per rack, and the generation after that could move toward 1 MW per rack in the same physical footprint. That is not normal load growth. That is a complete change in the power density model.

At that density, the old architecture starts to look strained. It is not because AC power suddenly became obsolete. It is because every conversion stage, every amp of current, every copper bus, every cable pathway, and every thermal assumption starts to matter more. When the rack density climbs by an order of magnitude, the supporting electrical infrastructure cannot simply be made a little larger and expected to behave the same way.

That is why the industry is paying attention to 800 VDC and solid-state transformers.

Not as science fiction. Not as a clean-tech side quest. As a practical response to the physics of AI infrastructure.

The problem is not just more power. It is where the power has to go.

The power problem in AI data centers is often discussed at the campus level. A project needs 100 MW, 300 MW, 1 GW, or more. That is the utility interconnection problem, and it is real. But the more interesting problem is what happens after the power gets on site.

Once power enters the facility, it has to move through a dense and increasingly unforgiving chain. Medium voltage is transformed down. Low-voltage AC is distributed. UPS systems condition and protect the load. PDUs distribute power closer to the rack. Server power supplies convert AC to DC because GPUs, CPUs, memory, and boards ultimately operate on DC internally.

That layered approach creates loss and complexity. Vertiv’s 2026 Frontiers report notes that many existing data centers still rely on hybrid AC/DC distribution from grid to IT racks, involving three to four conversion stages. Vertiv also argues that this approach faces challenges as AI-driven densities increase, especially as traditional 415 VAC or 480 VAC architectures run into limits around copper volume, thermal loss, and space efficiency.

This is where 800 VDC becomes important. Higher voltage allows the same amount of power to move at lower current. Lower current means less conductor mass, less heat, less voltage drop, and less space consumed by copper. When a facility is trying to feed megawatts into a smaller footprint, that is not a minor advantage. It becomes a design requirement.

Moonshot has already written about why 800 VDC is on the table, and the basic physics are straightforward: power equals voltage times current. If voltage rises, current can fall for the same power delivery. Since resistive losses scale with current squared, reducing current can have an outsized impact on heat and efficiency. Moonshot’s existing 800 VDC post also points out that even a 1% efficiency improvement in a 100 MW facility equals 1 MW of continuous savings, which becomes millions of kilowatt-hours over a year.

That is the part owners understand quickly. A fraction of a percent is no longer rounding error when the facility is operating at AI scale.

Solid-state transformers change the conversation

A traditional transformer does one primary job. It changes voltage through electromagnetic induction. It is big, heavy, proven, and familiar. It is also part of a broader system that may still require separate conversion and power conditioning equipment depending on the architecture.

A solid-state transformer, or SST, changes that relationship. In the pv magazine interview, SolarEdge describes an SST as not just replacing the transformer, but also replacing the converter. In their words, it functions more like a transformer and inverter together in one device. SolarEdge also describes using silicon-carbide components, similar to advanced inverter technology, to make the system smaller and more efficient than traditional transformer designs.

That is a major architectural shift.

Instead of taking medium voltage AC, stepping it down through a conventional transformer, then moving through additional conversion stages before finally reaching DC at the rack, an SST-based architecture points toward a more direct medium-voltage-to-DC path. SolarEdge says its SST introduces direct medium-voltage to DC architecture and expects 99% conversion efficiency from the in-development product.

Whether every claim becomes standard in the field remains to be seen. Early technologies always have commercialization hurdles. But the direction matters: the industry is trying to remove steps from the power chain.

Fewer steps means fewer losses. Fewer conversion points. Fewer heat sources. Fewer coordination boundaries. Potentially fewer pieces of large, custom-built gear sitting in a lead-time queue.

That last point is not trivial. Traditional transformers are often custom-built and can require orders years in advance. SolarEdge argues that SST manufacturing more closely resembles inverter manufacturing, using modular, mass-produced subcomponents that can support faster delivery. If that proves out at scale, it could change how data center teams think about long-lead electrical procurement.

The implication is bigger than one product category. It means data center power design may start borrowing more from power electronics, renewables, EV charging, industrial drives, and grid-interactive systems than from the old low-voltage building distribution playbook.

This does not mean AC disappears tomorrow

There is a temptation to turn every new architecture into a binary prediction. AC is dead. DC wins. Transformers go away. Everything changes overnight.

That is not how infrastructure works.

The data center industry is too large, too regulated, too risk-sensitive, and too capital-intensive to flip all at once. Existing facilities will still operate on established AC distribution. Traditional switchboards, transformers, UPS systems, PDUs, and protection schemes are not going away because a better future architecture exists. Standards, safety practices, available equipment, maintenance training, and inspector familiarity all matter.

Vertiv’s report takes a more realistic view. It says hybrid AC and DC systems are already common, and higher-voltage DC is likely to become more prevalent as rack densities increase. It also suggests AC may be used more selectively in the future, including at the grid interface, while DC becomes more important for internal power distribution.

That is the key point for owners and engineers: the transition will be hybrid.

Some facilities will adopt DC architecture in high-density AI zones first. Others will test it in modular pods or dedicated AI halls. Some will use traditional AC upstream and higher-voltage DC downstream. Some will wait for the standards, equipment ecosystem, and service workforce to mature before they commit.

That means the next several years are not about choosing between “old AC” and “new DC.” They are about designing infrastructure that can bridge both worlds without becoming obsolete halfway through its useful life.

That is where the switchboard conversation becomes more interesting.

Where legacy UL891 assumptions start to strain

UL891 switchboards are not the enemy of this transition. They remain central to a large portion of today’s low-voltage distribution reality. The mistake is assuming that the same low-voltage distribution assumptions will support every future density scenario without adjustment.

As rack densities climb, switchboard design has to account for higher service capacity, higher upstream fault current, tighter thermal margins, increased sustained load, and future voltage architecture flexibility. Moonshot’s existing 800 VDC post makes this point directly: even facilities not adopting full 800 VDC architectures will still feel the impact of higher-density power strategies, and UL891 assemblies must be evaluated with realistic fault studies and sustained load modeling.

This is where projects can get fooled. A switchboard may be perfectly appropriate for a traditional data center distribution system, but the future operating environment may not remain traditional. If the facility is designed today for 20 kW racks but expected to support 80 kW or 100 kW racks later, the electrical system has to be designed around that trajectory. If the site may eventually incorporate DC distribution, battery storage, on-site generation, or SST-based conversion, the early design choices need to preserve flexibility.

That does not mean every UL891 switchboard needs to be “800 VDC-ready,” whatever that phrase might mean in a given context. It means the upstream and downstream architecture needs to be planned intentionally. Cable routing, service entrance sizing, fault current assumptions, breaker strategy, compartment layout, heat dissipation, future tie-ins, metering, and documentation all become more important.

The worst outcome is not choosing AC today. The worst outcome is choosing an AC architecture today that blocks the facility from adapting tomorrow.

The timeline is closer than people think

The useful question is not, “Will every data center use SSTs next year?” They will not.

The better question is, “When do owners need to start designing around the possibility?”

The answer is now.

SolarEdge says it is aiming to be ready for the next generation of AI rigs that require 800 VDC, expected around 2027. Vertiv notes that higher-voltage DC architectures are already being developed, with deployments planned to support increasing GPU rack and pod power densities. Vertiv also recommends that early adopters investigate higher-voltage DC, particularly where on-site generation or microgrid deployment is part of the strategy.

That means owners making infrastructure decisions in 2026 are already making decisions that will either support or constrain the 2027 to 2030 hardware cycle.

This is especially important because electrical infrastructure has a longer life than AI hardware. Servers turn over quickly. GPU generations move fast. Power distribution equipment does not get replaced casually. A switchboard released today may still be expected to support much denser and more complex loads five years from now.

That gap between IT refresh cycles and electrical infrastructure cycles is where a lot of long-term risk lives. The compute layer changes every few years. The electrical layer is expected to last much longer. If the electrical layer is designed only for today’s GPUs, tomorrow’s business plan may have a problem.

What this means for Moonshot customers

The practical takeaway is not that every customer should immediately redesign around SSTs. That would be premature for many projects.

The practical takeaway is that every serious AI infrastructure project should start asking better power architecture questions.

If the facility is being built around high-density AI, the team needs to understand where AC distribution still makes sense, where higher-voltage DC could become relevant, and where the design should leave room for future conversion. If the site is evaluating on-site generation, battery energy storage, or microgrid capability, the DC conversation becomes even more important because solar, batteries, and fuel cells natively produce or store DC power. Vertiv specifically connects higher-voltage DC architectures with energy storage and possible DC microgrids integrating on-site generation and storage.

That is not a narrow equipment discussion. It is a full power strategy discussion.

Moonshot’s role is to help customers make that discussion practical. Moonshot’s UL891 switchboard page emphasizes that its engineering team designs solutions around supply chain constraints and flexibility, can use major breaker manufacturers, and evaluates site design during the sales process to provide code-compliant, cost-effective, faster solutions.

That is exactly the kind of thinking this transition requires.

Moonshot is not just building switchboards for current drawings. The opportunity is to help customers design distribution infrastructure that can handle today’s UL891 reality while preparing for a future where 800 VDC, SSTs, DC-compatible backup systems, DC microgrids, and high-density AI pods become more common.

That means asking questions like:

  • Does the current design assume rack density stays flat, or does it support meaningful load growth?
  • Are service entrance and downstream distribution choices flexible enough for future voltage architecture changes?
  • Are fault current studies, breaker selections, and bus ratings being modeled against future phases, not only phase one?
  • Will the facility be able to integrate batteries, on-site generation, or DC distribution without wholesale replacement?
  • Is the switchboard package being treated as a static procurement item, or as part of a long-term power architecture?

Those are the questions that separate a build that works on day one from one that survives the next generation of AI hardware.

The real takeaway

Solid-state transformers and 800 VDC are not just new electrical buzzwords. They are signs that the data center power stack is being rewritten from the rack backward.

AI has pushed compute density to the point where traditional distribution architecture is no longer the default answer for every case. The old AC chain can still work, but it becomes harder to scale cleanly when racks move from 10 kW to 100 kW and beyond. Each conversion stage, each conductor, each thermal margin, and each protection decision becomes more consequential.

SSTs point toward a future where power conversion is more integrated, more modular, and more directly connected to DC distribution. 800 VDC points toward a future where lower current and fewer conversion steps make extreme density more practical. Neither shift eliminates the need for disciplined engineering. If anything, they make early coordination more important.

For owners, the mistake would be waiting until the technology is fully mainstream before designing for it. By then, the facility may already be boxed in.

For Moonshot, the message is clear: power architecture is no longer just about sizing equipment. It is about building an adaptable path from today’s UL891 systems to tomorrow’s high-density AI infrastructure.

If you are planning an AI deployment, evaluating voltage architecture, or trying to understand how today’s distribution decisions will hold up over the next five years, early engineering alignment matters.

https://moonshotus.com/request-form/

Because the next generation of compute will not wait for yesterday’s power architecture to catch up.